Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

DNA methylation is a key epigenetic regulator of gene expression and a well-established hallmark of cancer. However, clinical methylation profiling remains costly and time-consuming, limiting routine use, particularly in low-resource settings. Inferring DNA methylation from routinely acquired histology images may serve as a scalable and cost-effective surrogate approach, providing early indications of methylation patterns prior to molecular profiling. We propose GraphCpG, a graph neural network-based framework for predicting continuous DNA methylation beta values across 10,000 CpG sites from H&E-stained whole slide images (WSIs) of gliomas, a tumour class in which methylation plays a central diagnostic and prognostic role. We evaluate the inferred methylation profiles in clinically relevant downstream tasks, including survival stratification and prediction of mitotic and immune subtypes, which capture tumour proliferative activity and tumour microenvironment states, respectively. Across TCGA glioma cohorts, predicted methylation features support significant survival separation and improved subtype classification performance relative to clinical and histology-only baselines, with AUROC gains of ≈ 17% and 3%, respectively. In some settings, methylation-based models achieve performance comparable to multi-modal approaches, and in certain cases yield improvements of up to ≈ 2% AUROC, highlighting the predictive value of histology-derived methylation signals. Pathway enrichment of well-predicted CpG-associated genes identifies immune and inflammatory processes consistent with glioma biology, providing supportive evidence for the biological relevance of the inferred signals. To the best of our knowledge, this is the first study to demonstrate that immune and mitotic subtypes can be predicted from DNA methylation profiles inferred directly from WSIs, highlighting the potential of scalable, image-based molecular characterization in glioma.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5013_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{RazMan_CrossModal_MICCAI2026,
        author = { Raza, Manahil AND Dawood, Muhammad AND Qaiser, Talha AND Rajpoot, Nasir},
        title = { { Cross-Modal Prediction of DNA Methylation from Histology Images } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors have, for the first time, demonstrated GraphCpG, a GNN based network that can predict DNA methylation beta values across 10000 CpG islands directly from H&E stained WSIs. This has led them to faciliatate the process of inferring clinical and molecular parameters like survival analysis, mitotic subtype prediction and immune subtype classification. As per their claim GraphCpG based methylation profiles were consistently outperformed both clinical and histology-only baselines, suggesting that immune subtype characteristics are more effectively captured through DNA methylation than through direct image-derived features.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The authors have demonstrated a novel methodology to capture CpG islands and methylation patterns from WSIs using GraphCpG GNN for DNA methylation pattern prediction. Very few studies have attempted to gauge the deep learning aspect of methylation beta values to predict molecular and clinical parameters. For example, Hoang et al, 2024 in Nature Medicine introduced a deep learning model, DEPLOY, that classifies CNS tumors from WSIs, generates predictions for DNA methylation beta values, and classifies tumor types into ten methylation classes in all CNS tumours using MLP (multilayer perceptron). Whereas in this case, the authors and team have built a novel methodology by using GNN, primarly on gliomas TCGA dataset and succesfully predicted survival analysis, mitotic subtypes and immune checkpoints after regression analysis based on methylation patterns. Now, this manuscript has used tissue segmentation, path embedding using Delaunay triangulation for better histological assesment by using spatial graphs for better geometric alignment. Furthermore, downstream analysis of the immunological aspects using Pathway Enrichment performed using the MSigDB hallmark 2020 library with Enrichr, and then confirmed by GSEA, showed correlation of TNF-α/NF-κB signaling, Coagulation and Apoptosis, NF-κB signaling drives proliferation, inflammation with relevant methylation patterns linked to immune and inflammatory processes in gliomas. The prediction utility was further validated by comparing GraphCpG with a comprehensive experimental framework (uni-modal, multi-modal, and clinical baselines) which proves as an excellent evidence for clinical utility.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    All evaluations rely on internal 3-fold cross-validation within TCGA, given subclass heterogeneity and interobserver variability across institutions, scanners, and staining protocols, external validation cohort is essential. One of my primary concerns with the paper is that even though it projects the methodology and the technical aspect very logistically however working on 450k epic array data is redundant, theoritically it is ok otherwise it does not make much sense because we have 935k or 850k Illumina Epic Chip array data available in the datasets now, so for practically imprelemting it, its essential to use the updated CpG island dataset. Only on 502 patients’ data isnt a strong sample size but to begin with its not so bad. The main issue is that gliomas have molecular/intratumoral heterogeneity. The analysis is blurred unless predictions are derived from true GBM or LGG cases, hence lack of transparency of sublasses used for methylation pattern prediction will be misleading practically. The authors mention that “Methylation-based models consistently outperformed both clinical and histology-only baselines, suggesting that immune subtype characteristics are more effectively captured through DNA methylation than through direct image-derived features” is a bold claim and therefore needs validation in a larger and external cohort to prove this statement further. Median Spearman 0.42, only 42% variance, insufficient for confident clinical deployment without improving regression performance.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper justifies strong computational ground and adheres to CAI aspects but falls short in the clinically translational aspect. The novel approach to predicting methylation pattern if expanded and verified using a robust external cohort while ensuring the accurate inclusion of subclasses of gliomas, will enable better utilization of the GraphCpG model and also validate its robustness as a strong generazible protocol.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    Although the study is not yet ready for clinical translation, it presents a technically novel graph-based framework for continuous methylation inference from histology and demonstrates consistent downstream utility across survival and subtype prediction tasks. The rebuttal clarifies the main methodological concerns, including within-fold feature selection and model specification, and the authors now appropriately frame external validation as future work. For these reasons, I consider the paper a solid computational pathology contribution and recommend acceptance, while encouraging a more ambitious external validation study in future work.



Review #2

  • Please describe the contribution of the paper

    The paper proposes GraphCpG, a graph neural network that predicts continuous DNA methylation beta values from H&E whole slide images in glioma. The predicted profiles are then used for survival analysis and for classifying mitotic and immune subtypes in matched TCGA LGG and GBM cases.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The clinical motivation is real. Molecular profiling is slow and uneven in practice, and a histology-based surrogate for methylation-derived signals would matter if it works.
    • Histology-to-methylation prediction isn’t new, but pushing to continuous CpG-level values at this scale, and then asking whether the inferred profiles are actually useful for glioma-relevant tasks, is the right next step.
    • The graph formulation is a relatively novel choice in this setting, and the motivation — capturing spatial tissue architecture rather than treating patches as a bag — is sound.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The training objective for the multi-output prediction is the single biggest thing I couldn’t resolve on my own. I spent a while trying to reconcile Equations 2-4 with the 10,000 CpG claim and couldn’t land on a consistent reading. Either z_i is a vector and Equation 4 needs re-expressing, or the task is not what Section 2.1 describes.
    • The correlation-based CpG filtering and top-1k selection need to be fold-internal. As written, I can’t tell whether they are, and if they aren’t, the downstream numbers are optimistic.
    • The survival analysis is hard to read without clinical-only and histology-only baselines. On its own, the reported number doesn’t tell me how much the methylation prediction is adding.
    • The immune subtype result is the one I’d push hardest on. LGG and GBM have very different immune landscapes, and combining them means the classifier may be partly learning tumor type or grade. This needs to be acknowledged directly and reflected in the interpretation of the result.
    • The sample counts don’t add up: 502 patients in Section 2.1, but 417 cold + 124 hot = 541 in Section 3.1.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    1.The dimensionality of z_i^(l), g^(l), and F(G) in Section 2.4 is the first thing I’d fix. From the text, z_i^(l) reads as a patch-level prediction and F(G) as slide-level, but it’s never stated whether these are scalars or 10,000-dim vectors. Equation 4 is written as a scalar pairwise ranking loss, which doesn’t square with a 10,000-output task. If the ranking loss is applied per-CpG and then aggregated, say so; if something else is happening, the equation needs to reflect that. 2.State explicitly that correlation-based filtering and top-k CpG selection are performed on the training fold only. If not, that should be acknowledged clearly and the downstream results interpreted accordingly. 3.The sample count discrepancy in Section 3.1 needs to be resolved — 502 vs. 541 isn’t a rounding issue. 4.The survival claim is stronger than the evidence supports without clinical-only and histology-only baselines. Without those comparisons, the framing should be tempered accordingly. 5.On the immune subtype task: please discuss the LGG/GBM confound directly. At minimum, the paper should acknowledge that tumor type may be driving part of the signal and interpret this result accordingly. 6.Table 1 — the fold variance eats most of the gaps you’re claiming as improvements. Several of these differences read as ties given the reported variability, and the text should reflect that. 7.The enrichment analysis is interesting but post hoc — it’s conditioned on the well-predicted CpGs, so it can support biological plausibility but can’t establish that the model has learned meaningful methylation structure. Framing it as suggestive rather than confirmatory would help. 8.Minor: Figure 2 labels are too small to read. 9.From a reproducibility perspective, please specify whether the CTransPath encoder is frozen or fine-tuned, and clarify the exact implementation of the multi-output loss. Given the number of moving parts in the pipeline, a code release would be materially helpful.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This is a promising paper on a problem I think matters, and I like the direction — the graph formulation and the move toward clinically relevant glioma tasks are both the right instincts. My score is borderline because several methodological things I’d need to interpret the results aren’t resolvable from the text: the multi-output objective, the fold structure of the downstream selection, and the sample-count inconsistency. These are all fixable in rebuttal, and if they are, my concerns would drop substantially.

  • Reviewer confidence

    Somewhat confident (2)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I maintain an accept recommendation, but this remains a borderline paper. I do not view the lack of external validation or definitive translational impact as sufficient grounds for rejection in a MICCAI methodological submission, but the final version needs much more careful reporting and interpretation.

    The rebuttal directly addressed two of my most important concerns. First, the authors clarified that the patch-level predictions, pooled layer outputs, and slide-level prediction are all C-dimensional vectors over the 10,000 CpG sites, and that the ranking loss is applied per CpG and aggregated across CpG outputs. This resolves the dimensionality ambiguity in the submitted equations, although the manuscript still needs the corrected loss expression. Second, they confirmed that correlation-based CpG filtering and top-k CpG selection were performed within each training fold. This substantially reduces my concern that the downstream classification and survival numbers were inflated by test-fold leakage.

    Several concerns remain after the rebuttal. For the survival analysis, clinical-only and histology-only survival baselines are not part of the submitted evidence, so the current result should be framed as showing prognostic association of inferred methylation features rather than added value over simpler clinical or image-only models. I do not expect new survival experiments in rebuttal, but the claim should be narrowed accordingly. The immune subtype experiment remains potentially confounded by LGG/GBM composition; the rebuttal acknowledges this, but because no controlled or stratified analysis is in the submitted evidence, the immune-subtype claim should be interpreted cautiously. The sample-count discrepancy was acknowledged as corrected, but the rebuttal does not state the corrected counts, which leaves a reproducibility issue. The loss clarification also does not fully address calibration: a per-CpG ranking objective can support relative methylation ordering, but the paper should be careful when describing the outputs as continuous beta-value estimates. Finally, some reported gains in Table 1 are small relative to fold variability, and the rebuttal does not substantially change that interpretation.

    Overall, I think the paper is promising enough for acceptance because the central methodological direction is useful, the graph-based WSI formulation is reasonable, and the two most serious fixable issues from my review – the multi-output loss ambiguity and possible fold leakage in CpG selection – were answered credibly in the rebuttal. My acceptance is contingent in spirit on the final version explicitly correcting the loss formulation, reporting the corrected sample counts, stating the frozen encoder setting, and tempering the survival, immune-subtype, and enrichment-analysis claims.



Review #3

  • Please describe the contribution of the paper

    A method based on GNN was proposed to predict methylation values from pathological images, and the performance of downstream tasks was enhanced based on the predicted methylation information.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    (1) This work try to demonstrate that DNA methylation profiles predicted from WSI images can be used to classify subtypes in gliomas. This goes beyond prior work that only predicted methylation or tumour types, showing that histology-derived epigenetic signals capture clinically meaningful tumour phenotypes. (2) The authors evaluate predicted methylation on three clinically relevant tasks: survival stratification, mitotic subtype classification, and immune subtype classification. They compare against clinical, histology‑only, methylation‑only, and multi‑modal baselines. The methylation‑only model outperforms histology and clinical baselines, and in some settings matches multi‑modal approaches, demonstrating the practical utility of the inferred profiles.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    (1) Limited novelty of predicting methylation from histology. Some previous works already predict continuous DNA methylation values from H&E images, including for LGG and GBM. The proposed GNN offers incremental architectural changes but is not shown to outperform existing methods, as no direct comparison is provided (Section 3.1 Regression Results). (2) The accuracy of the model is not high, and the selection criteria are not clear. Firstly, for the prediction of methylation values, the correlation coefficient is set at a threshold of 0.4, which itself is not very strong. And in subsequent tasks, sometimes the top 10 are selected, such as survival analysis (Section 3.1), and sometimes the top 1000 are selected, such as classification prediction tasks (Table 1). This highly selective pipeline may overestimate clinical utility.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The innovation of the method is insufficient. It is suggested that this work be further expanded into a clinical application under the framework of two-stage prediction, which will be interesting.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    I still have concerns about the methodological innovation of this work. If only GNN is used for modeling, it is equivalent to solving a new task using traditional methods, and this work is not a pioneering work based on HE image prediction of methylation values. I suggest that the author should consider how to better design a specific network architecture to handle methylation prediction tasks, in order to solve the practical dilemma of insufficient prediction performance.



Author Feedback

Rebuttal to the Reviewers’ Comments General Remarks: First, we thank the reviewers for their insightful comments and careful observations. We note that all three reviewers recognized the importance of our work for predicting continuous CpG methylation profiles from H&E histology images using graph-based WSI formulation and the value of evaluating these inferred methylation signals across downstream glioma-specific clinically relevant tasks. In the remainder of this rebuttal, Rx refers to Reviewer# x and MRx refers to Meta Reviewer’s point# x.

[MR1/R2] Prediction Setup and the Ranking Loss: The multi-output formulation is in Section 2.4.In the revised paper, we will explicitly state that the predictions z_i^(l), g^(l), and F(G) are all C-dimensional vectors corresponding to predictions across all C=10,000 CpG sites. We will also revise Eq.(4) to explicitly define the ranking loss for each CpG site and aggregate the losses across CpG outputs.

[MR2/4/R2] Clarity: We confirm that correlation-based CpG filtering and top-k CpG selection were performed within each training fold. We will revise the discussion surrounding the enrichment and survival analyses to present these findings more cautiously and as supportive rather than confirmatory evidence of biological relevance. We have corrected the discrepancy in the reported sample counts in Section 3.1 and clarify that the CTransPath encoder was frozen during feature extraction.

[MR3] Performance Interpretation and External Validation: The reported median correlation (0.42) should be interpreted in the context of cross-modal histology-to-methylation prediction tasks, where moderate correlations (0.27, 0.33[1], 0.45[2]) have been reported due to substantial biological and technical variability. We agree that external validation would further strengthen the findings. Accordingly, we will revise the paper to add this point to the discussion as part of future work.

[MR4/R3] Positioning Relative to Prior Work: Prior approaches primarily employed MLP-based or conventional ML methods for continuous methylation prediction, while GraphCpG uses a GNN to model spatial relationships between histological regions for predicting methylation profiles across 10,000 CpG sites. Feature Selection: The threshold of 0.4 follows prior methylation prediction studies [1, 2] and was used to identify comparatively well-predicted CpG sites. The downstream feature-selection procedures were task-specific: smaller CpG subsets were used for Cox regression, whereas larger subsets were retained for classification tasks using MLP classifiers. We included evaluations using all CpG sites, correlation-filtered CpGs and top-1,000 CpG subsets for classification tasks to assess the effect of feature selection on downstream performance.

[MR4] Immune Subtype Analysis: We acknowledge that differences between tumor classes may contribute to part of the predictive signal in addition to immune subtype characteristics, and we will interpret the findings accordingly.

[R1] Methylation Platform: Regarding methylation platforms, the TCGA LGG and GBM datasets predominantly use HM450 arrays, although we agree that EPIC/850k arrays are more ubiquitous nowadays. We will explicitly state this as a limitation of our work, if the paper is accepted.

[R3] Novelty Challenge: As the MR has established, the reviewers recognize several strengths, chiefly: (a) the graph-based formulation for prediction of continuous CpG-level methylation values, and (b) the evaluation on clinically relevant downstream tasks in survival analysis and prediction of mitotic subtypes and immune subtypes.

[1] Hoang, Danh-Tai, et al. “Path2Omics Enhances Transcriptomic and Methylation Prediction Accuracy from Tumor Histopathology.” Cancer research 85.24 (2025): 5098-5112.[2] Hoang, Danh-Tai, et al. “Prediction of DNA methylation-based tumor types from histopathology in central nervous system tumors with deep learning.” Nature Medicine 30.7 (2024): 1952-1961.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The reviews are mixed, with one weak-accept, one weak-reject, and one reject recommendation. The paper addresses an important and clinically relevant problem: predicting DNA methylation profiles from routine H&E whole-slide images and using the inferred profiles for glioma-related downstream tasks. The reviewers recognize several strengths, including the graph-based WSI formulation, the attempt to predict continuous CpG-level methylation values, and the evaluation on survival, mitotic subtype, and immune subtype prediction.

    However, the reviewers raise substantive concerns. The main issues are: (1) the multi-output methylation prediction objective is not sufficiently clear, especially how the ranking loss and slide-level predictions apply to 10,000 CpG outputs; (2) the CpG filtering and top-k selection procedures must be clearly shown to be fold-internal to avoid optimistic downstream results; (3) the evidence for clinical utility is limited by internal TCGA cross-validation only, lack of external validation, and modest methylation regression performance; and (4) several interpretation issues need clarification, including the LGG/GBM confounding in immune subtype classification, missing clinical/histology baselines for survival analysis, lack of direct comparison with prior histology-to-methylation methods, and the reported sample-count inconsistency.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    After considering the reviews and rebuttal, I recommend acceptance. The paper addresses an important computational pathology problem by predicting continuous CpG-level methylation profiles from H&E whole-slide images and evaluating the inferred profiles on glioma-related downstream tasks. The rebuttal resolves two important technical concerns: the authors clarified that the methylation predictions are C-dimensional CpG vectors and that the ranking loss is applied per CpG and aggregated, and they confirmed that CpG filtering and top-k selection were performed within each training fold. These clarifications reduce concerns about objective ambiguity and potential test-fold leakage.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    After reviewing the paper in detail, I share the concerns regarding the level of methodological innovation in this work. The paper would be substantially strengthened by a more task-specific network architecture tailored to the unique characteristics of methylation prediction, particularly to better address the practical challenge of limited predictive performance. Given the current level of methodological novelty and the limited empirical support, this was not an easy decision, I regretfully have to recommend rejection.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The work lacks specific architectural design and algorithmic innovation, remaining an application of existing models rather than a methodological contribution.



back to top